Snowgander: AI Vendor Abstraction Layer

---

Check out Snowgoose!

Snowgander is the core abstraction engine powering Snowgoose, an intelligent agent framework.

If you're building AI-powered applications, agents, or workflows, Snowgoose provides the tools and structure you need on top of Snowgander.

Explore the Snowgoose App: snowgoose.app Check out the Snowgoose Repo: github.com/loneyeti/snowgoose

---

What is Snowgander?

Snowgander is a TypeScript library that makes it easy to talk to different AI models (like those from OpenAI, Anthropic, Google, and OpenRouter) using one simple, consistent interface.

Stop writing vendor-specific code! Use Snowgander to:

• Easily switch between AI models and vendors.
• Use a single method (sendChat) for most conversational interactions.
• Stream responses in real-time for interactive UIs.
• Handle different types of content (text, images, tool use, thinking steps) uniformly.

How to Use It: The Basics

1. Installation

npm install snowgander

2. Configuration (Set API Keys)

Do this once when your application starts. The factory holds the configuration for each vendor.

import { AIVendorFactory } from "snowgander";

// Example: Load keys from environment variables
if (process.env.OPENAI_API_KEY) {
  const config = { apiKey: process.env.OPENAI_API_KEY };
  AIVendorFactory.setVendorConfig("openai", config);
  // Also configure 'openai-image' if using DALL-E via OpenAIImageAdapter
  AIVendorFactory.setVendorConfig("openai-image", config);
}

if (process.env.ANTHROPIC_API_KEY) {
  AIVendorFactory.setVendorConfig("anthropic", {
    apiKey: process.env.ANTHROPIC_API_KEY,
  });
}

if (process.env.GOOGLE_API_KEY) {
  AIVendorFactory.setVendorConfig("google", {
    apiKey: process.env.GOOGLE_API_KEY,
  });
}

if (process.env.OPENROUTER_API_KEY) {
  AIVendorFactory.setVendorConfig("openrouter", {
    apiKey: process.env.OPENROUTER_API_KEY,
    // OpenRouter's baseURL is the default, but you can override it
    baseURL: "https://openrouter.ai/api/v1",
  });
}

3. How to Initiate an Adapter

To get an adapter, you need the vendor name (e.g., "openai") and a ModelConfig object describing the specific model you want to use.

import { AIVendorFactory, ModelConfig, AIVendorAdapter } from "snowgander";

// --- Prepare ModelConfig (Example) ---
// In a real app, you'd likely fetch this from a database or config file.
const gpt4oConfig: ModelConfig = {
  apiName: "gpt-4o", // The name the vendor API expects
  isVision: true, // Can it process images?
  isImageGeneration: true, // Can it generate images?
  isThinking: false, // Does it support structured thinking output?
  inputTokenCost: 5, // Cost per million input tokens (optional)
  outputTokenCost: 15, // Cost per million output tokens (optional)
  imageOutputTokenCost: 20, // Special cost for image generation output
  webSearchCost: 0.05, // Flat fee when a web search tool is used
};

// --- Get the Adapter ---
let adapter: AIVendorAdapter;
try {
  const vendorName = "openai"; // Or "anthropic", "google", etc.
  adapter = AIVendorFactory.getAdapter(
    vendorName,
    gpt4oConfig // Pass the config for the specific model
  );

  console.log(
    `Successfully got adapter for ${vendorName} / ${gpt4oConfig.apiName}`
  );
  // Now you can use the 'adapter' instance!
} catch (error) {
  console.error("Failed to get adapter:", error);
}

4. Making API Calls

Snowgander provides two primary ways to interact with AI models: a standard request-response model and a real-time streaming model.

A. Standard Request/Response (generateResponse)

The generateResponse method is the fundamental way to get a complete response from an AI. You provide all the messages at once, and it returns the full response after the model has finished processing.

import { AIRequestOptions, AIResponse, TextBlock } from "snowgander";

// Assume 'adapter' is an AIVendorAdapter instance obtained above

async function getStandardResponse() {
  const options: AIRequestOptions = {
    model: gpt4oConfig.apiName,
    messages: [
      {
        role: "user",
        content: [
          { type: "text", text: "Tell me a short joke about computers." },
        ],
      },
    ],
    systemPrompt: "You are a witty comedian.",
    maxTokens: 100,
  };

  try {
    console.log("Sending request to AI...");
    const response: AIResponse = await adapter.generateResponse(options);
    console.log("AI Response Received!");

    // Process the response content (see section 5 for details)
    const assistantMessage = response.content
      .filter((block): block is TextBlock => block.type === "text")
      .map((block) => block.text)
      .join("");

    console.log("Assistant:", assistantMessage);

    if (response.usage) {
      console.log(`Cost: $${response.usage.totalCost.toFixed(6)}`);
    }
  } catch (error) {
    console.error("Error generating response:", error);
  }
}

getStandardResponse();

B. Streaming Responses (streamResponse)

For interactive applications where you want to display the response as it's being generated (like a "typewriter" effect), use the streamResponse method. It returns an AsyncGenerator that yields response chunks (ContentBlocks) in real-time.

import { AIRequestOptions, ContentBlock, TextBlock } from "snowgander";

// Assume 'adapter' is an AIVendorAdapter instance obtained above

async function getStreamingResponse() {
  const options: AIRequestOptions = {
    model: gpt4oConfig.apiName,
    messages: [
      {
        role: "user",
        content: [{ type: "text", text: "Write a short poem about the sea." }],
      },
    ],
    systemPrompt: "You are a thoughtful poet.",
    maxTokens: 150,
  };

  try {
    console.log("Streaming response from AI...");
    const stream = adapter.streamResponse(options);
    if (!stream) {
      console.log("This adapter does not support streaming.");
      return;
    }

    process.stdout.write("Assistant: ");
    // Use a for-await-of loop to process chunks as they arrive
    for await (const chunk of stream) {
      if (chunk.type === "text") {
        // In a real app, you would append this to your UI
        process.stdout.write(chunk.text);
      } else if (chunk.type === "error") {
        console.error(`\nStream Error: ${chunk.privateMessage}`);
        break;
      }
      // You can also handle other chunk types like 'thinking' or 'image_data'
    }
    console.log("\n--- End of Stream ---");
    // Note: Final usage/cost information is not available when streaming.
  } catch (error) {
    console.error("\nError streaming response:", error);
  }
}

getStreamingResponse();

C. Stateful Conversations (sendChat)

For building chatbots, the sendChat method is a convenient helper. It wraps generateResponse and simplifies state management. You pass it a Chat object containing the full conversation history and the new prompt, and it internally constructs the request for you.

import { Chat, ChatResponse, TextBlock } from "snowgander";

// Assume 'adapter' is an AIVendorAdapter instance obtained above

// --- Manage this Chat object state in your application ---
let currentChat: Chat = {
  model: gpt4oConfig.apiName,
  responseHistory: [], // Start with an empty history for a new conversation
  prompt: "Tell me a short joke about computers.", // The user's first message
  systemPrompt: "You are a witty comedian.",
  maxTokens: 100,
  visionUrl: null,
  budgetTokens: null,
};

async function haveAConversation() {
  try {
    console.log(`User: ${currentChat.prompt}`);
    const response: ChatResponse = await adapter.sendChat(currentChat);
    console.log("AI Response Received!");

    const assistantMessage = response.content
      .filter((block): block is TextBlock => block.type === "text")
      .map((block) => block.text)
      .join("");
    console.log("Assistant:", assistantMessage);

    // --- IMPORTANT: Update your chat state for the next turn ---
    // 1. Add the user's prompt to the history
    currentChat.responseHistory.push({
      role: "user",
      content: [{ type: "text", text: currentChat.prompt }],
    });
    // 2. Add the AI's full response to the history
    currentChat.responseHistory.push(response);
    // 3. Clear the prompt, ready for the next user input
    currentChat.prompt = "";
  } catch (error) {
    console.error("Error sending chat:", error);
  }
}

haveAConversation();

5. Understanding the Response

Both generateResponse and sendChat return a response object (AIResponse or ChatResponse) with this structure:

interface ChatResponse {
  role: string; // Usually "assistant" (or "error" if something went wrong)
  content: ContentBlock[]; // An array containing the AI's response parts
  usage?: {
    // Optional cost information
    inputCost: number;
    outputCost: number;
    totalCost: number;
    // ... other optional cost fields
  };
}

• role: Identifies the sender (usually "assistant").
• content: This is an array (ContentBlock[]). Each item in the array represents a part of the response. Common types include:
  • TextBlock: Contains plain text ({ type: 'text', text: '...' }).
  • ImageDataBlock: Contains generated image data ({ type: 'image_data', ... }).
  • ThinkingBlock: Contains structured thinking steps from the model ({ type: 'thinking', ... }).
  • ToolUseBlock: Indicates the AI wants to use a tool that your application must handle.
  • ErrorBlock: If an error occurred ({ type: 'error', ... }).
• usage: If available, provides the estimated cost for the interaction based on token counts.

Your application needs to:

1. Check the role (e.g., handle an "error" role).
2. Iterate through the content array to process and display the different blocks appropriately (e.g., render text, display images, trigger tool execution).
3. For conversations, add the entire response object to your responseHistory so the AI has context for the next turn.

Anthropic Models: Thinking, Effort, and Sampling Parameters

Snowgander stays a thin wrapper: to use a new Claude model (Opus 4.6/4.7/4.8, Sonnet 4.6/5, Fable 5, Mythos 5), just set apiName to the model string in your ModelConfig — no code changes to generateResponse, streamResponse, or sendChat are required. The Anthropic adapter detects the right request shape automatically based on the model string.

⚠️ One behavior change to be aware of when adopting the newest models: claude-opus-4-7, claude-opus-4-8, claude-sonnet-5, claude-fable-5, and claude-mythos-5 reject temperature/topP entirely — the adapter throws if either is set, not just both together. If your existing code always sets a default temperature, remove it (or branch on model) before pointing that same call at one of these models, or generateResponse/sendChat will throw instead of silently ignoring it.

Models	temperature / topP	thinkingMode behavior	effort levels
Claude 3.x	Both allowed together	budgetTokens → budget_tokens	n/a
Opus 4.0–4.5, Sonnet 4.0/4.5, Haiku 3.x/4.5	At most one	budgetTokens → budget_tokens	n/a
Opus 4.6, Sonnet 4.6	At most one	Adaptive thinking	low / medium / high / max
Opus 4.7/4.8, Sonnet 5, Fable 5, Mythos 5	Neither allowed	Adaptive thinking only	low / medium / high / xhigh / max

For models that support adaptive thinking, budgetTokens is still accepted for backward compatibility and is automatically mapped to a reasonable effort level — or pass effort explicitly for direct control (including the newer "xhigh"/"max" levels):

const options: AIRequestOptions = {
  model: "claude-opus-4-8",
  messages: [{ role: "user", content: [{ type: "text", text: "Solve this." }] }],
  thinkingMode: true,
  effort: "xhigh", // low | medium | high | xhigh | max
  maxTokens: 4096,
  // Do NOT set temperature/topP here — Opus 4.8 will throw if either is set.
};

Chat/sendChat also accept these as optional fields (new — previously only generateResponse exposed them), so existing Chat objects that omit them are unaffected:

let currentChat: Chat = {
  model: "claude-opus-4-8",
  responseHistory: [],
  prompt: "Solve this.",
  maxTokens: 4096,
  visionUrl: null,
  budgetTokens: null,
  effort: "high", // optional — new
  outputFormat: undefined, // optional — new, structured output via output_config.format
};

Development

• Build: npm run build (Compiles TS to dist/)
• Test: npm test (Runs Jest tests)

---

Ready to build more complex AI agents?

Snowgander provides the vendor abstraction. Snowgoose provides the framework.

Give your AI projects structure and power with Snowgoose!

Explore the Snowgoose App: snowgoose.app Check out the Snowgoose Repo: github.com/loneyeti/snowgoose